Industries / Finance and business services / Staffing and recruiting
Screening and interviews
Deciding which applicants go forward: written tests, interviews, and the software that sorts and scores applications before a person reads them.
6
Milestones achieved, from 1883 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1883
Exams for federal jobs
After President James Garfield was killed by a disgruntled job seeker, the Pendleton Act of 16 January 1883 required covered US federal jobs to be filled through open, competitive examinations that were "practical in their character", instead of by political favor. It set up the Civil Service Commission to enforce the law.
Credit
Senator George Hunt Pendleton; President Chester A. Arthur, who signed it; US Congress
1917
Group tests for recruits
Between September 1917 and January 1919 the US Army gave written intelligence tests to 1,726,966 men in 35 training camps: the Alpha test for men who could read English, and a separate picture-based Beta test for those who could not. Examiners used the grades to recommend 7,800 men (0.5%) for discharge and others for labor or development battalions.
LessonA written test also measures reading, so the Army needed a second test to rank men who could not read English.
Credit
Robert M. Yerkes and the Army's Division of Psychology
1998
A ranking of methods
Frank Schmidt and John Hunter pooled 85 years of studies on how well each hiring method predicts job performance, measured as a correlation where 1 would be perfect prediction. They put work sample tests at 0.54 and general mental ability tests and structured interviews (the same planned questions and scoring for every applicant) at 0.51, and listed what each method added to an ability test. By 2023 the paper had been cited more than 6,500 times, by Sackett and colleagues' count.
Credit
Frank L. Schmidt and John E. Hunter (Psychological Bulletin, 1998)
2021
Video interviews drop faces
HireVue sells one-way video interviews, in which applicants record answers that software scores. In January 2021 it said it had removed facial analysis from its assessments, a change it dated to early 2020, because its own research found that advances in language analysis meant "visual analysis no longer significantly added value". The announcement came a year after the Electronic Privacy Information Center (EPIC) filed a complaint with the US Federal Trade Commission calling the tools unproven; HireVue kept analyzing speech and intonation.
LessonA signal that added little to prediction cost the vendor public trust, and HireVue said it "wasn't worth the concern".
Credit
HireVue; EPIC
2022
Validity estimates cut
Paul Sackett, Charlene Zhang, Christopher Berry and Filip Lievens showed that earlier summaries had over-corrected for range restriction, the fact that job performance can only be measured for the people who were hired. Recomputed, most methods fell by 0.10 to 0.20: structured interviews came first at 0.42, job knowledge tests followed at 0.40, work samples fell to 0.33 and general mental ability tests to 0.31.
LessonThe best-known ranking of hiring methods overstated how well all of them work, and ability tests lost their lead to structured interviews.
Credit
Sackett, Zhang, Berry and Lievens (Journal of Applied Psychology, 2022)
2025
AI writing blurs applications
Anaïs Galdin and Jesse Silbert studied about 2.7 million applications for coding jobs on Freelancer.com, an online marketplace for freelance work, which added an AI tool that writes applications in April 2023. Using a language model to score how closely each application was tailored to its job post, they found that before ChatGPT, employers paid more for workers who sent tailored applications, and afterward they did not. Their model estimates that if written applications carry no signal, the most able fifth of workers are hired 19% less often and the least able fifth 14% more often.
LessonWhen writing costs nothing, a tailored application stops telling employers who is good.
Credit
Anaïs Galdin (Dartmouth College, Tuck School of Business) and Jesse Silbert (Princeton University), working paper, November 2025
Watch for
AI voice interviews tested
In a pre-registered field experiment with 70,884 applications for customer service jobs in the Philippines, from March to June 2025, applicants were randomly assigned to a human recruiter or an AI voice agent following the same interview guide, and human recruiters made every hiring decision. The authors report that applicants interviewed by the AI were 12% more likely to receive a job offer, and that more of them started work and stayed, with no drop in measured productivity; given a choice, 78% chose the AI. The study is a working paper run with PSG Global Solutions, a recruitment firm owned by Teleperformance, and one author took an unpaid role at the firm after data collection ended.
Achieved whenthe study passes peer review, or a team with no tie to the firm repeats it and finds a similar effect.
Claimed by
Brian Jabarian (Carnegie Mellon University) and Luca Henkel (Erasmus University Rotterdam), July 2026
One in four profiles fake
Gartner predicts that by 2028 one in four candidate profiles worldwide will be fake. In its own survey of 3,000 job candidates in the second quarter of 2025, 6% admitted to interview fraud, either posing as someone else or having someone else pose as them.
Achieved whena study by a party outside Gartner, across many employers, measures the share of fake profiles or impostor interviews and finds it near that level.
Claimed by
Gartner, July 2025
Voice interviewer on validated rubrics
HireVue launched an AI Interviewer that holds two-way voice interviews with every applicant at any hour and scores the answers against rubrics it says are built from "thousands of outcome studies". The launch gives no accuracy figures for the new product.
Achieved whenan independent study reports how well the AI Interviewer's scores predict the later job performance of people hired, with sample sizes, and compares them with a structured interview run by trained people.
Claimed by
HireVue, June 2026
Open
01
Résumé parser scorecard
A public scorecard for the dedicated résumé parsers that hiring software uses to turn a résumé file into fields such as name, jobs, dates and skills. Run them on synthetic résumés with known answers, in several languages and layouts, including two-column designs and scanned pages.
The briefCredit · unclaimed02
Deepfake interview test
A test in which consenting actors take part in recorded video interviews, some of them wearing a live face swap or a converted voice. Score the detection tools sold to employers, and a panel of human interviewers, on telling real from swapped, with and without short challenges such as turning the head or passing a hand in front of the face.
The briefCredit · unclaimed
Open 01 · the brief
Résumé parser scorecard
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard for the dedicated résumé parsers that hiring software uses to turn a résumé file into fields such as name, jobs, dates and skills. Run them on synthetic résumés with known answers, in several languages and layouts, including two-column designs and scanned pages.
Why it is still open
Two public benchmarks exist, and neither tests these parsers. ResumeBench (Zijian Ling and colleagues, EMNLP 2025) scored 24 large language models on 2,500 synthetic résumés in five languages. Careerflow's ResumeBench leaderboard (November 2026) scores 18 general document-extraction services and AI models on 300 résumés; no dedicated résumé parser is on it. Parser vendors' accuracy figures are their own.
Built with what exists
The synthetic EMNLP ResumeBench set and its scoring method (licensed CC BY-NC 4.0, for non-commercial research); trial accounts or APIs of commercial résumé parsers; the open-source OpenResume parser; a free trial of an applicant tracking system, to see what survives import; Python.
First step
Run 50 English résumés from the EMNLP ResumeBench set through OpenResume and one commercial parser's trial, and score names, job titles and dates.
Done looks like
Field-level precision and recall for at least five parsers on at least 500 synthetic résumés across five languages and four layouts, published with code and each parser's raw output, with every vendor invited to respond before publication.
Open 02 · the brief
Deepfake interview test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test in which consenting actors take part in recorded video interviews, some of them wearing a live face swap or a converted voice. Score the detection tools sold to employers, and a panel of human interviewers, on telling real from swapped, with and without short challenges such as turning the head or passing a hand in front of the face.
Why it is still open
The FBI warned in June 2022 of applicants using voice spoofing, and possibly deepfakes, in remote interviews, with lip movements that did not match the audio. The independent tests found use recorded media: a Purdue University benchmark (October 2025) scored detectors from academia, government and industry on political deepfakes shared on social media, and NIST's GenAI: Deepfakes 2026 evaluation tests forensic detection on prepared images and video. GOTCHA, by Govind Mittal, Chinmay Hegde and Nasir Memon (first posted October 2022), scored challenge prompts against live face swaps on its own recordings, not commercial detectors in an interview. An August 2026 market report by Biometric Update and Goode Intelligence notes "a lack of benchmarking and independent evaluation data" for vendors' claims. No test of detectors inside live interview calls was found.
Built with what exists
Deep-Live-Cam, an open-source real-time face-swap tool; an open-source voice-conversion model; Zoom or Microsoft Teams; trial access to commercial detectors; GOTCHA's list of challenges; a written consent form for every actor.
First step
Record ten five-minute interviews with two consenting actors, five of them with Deep-Live-Cam running, and ask three colleagues to say which were swapped.
Done looks like
At least 100 recorded interview sessions, half of them swapped, with the detection rate and false-alarm rate for each detector and for a panel of at least 10 interviewers, published with the method, and with swapped clips watermarked or withheld.